test_unpooling.lua

local grad = require 'autograd'
local nn = require 'nn'

local W1 = torch.FloatTensor(1, 1, 10, 10):normal()
local pooler = nn.SpatialMaxPooling(2, 2, 2, 2)
local unpooler = nn.SpatialMaxUnpooling(pooler)

-- not sure why these lines are necessary.
pooler.indices = torch.FloatTensor()
pooler.output = torch.FloatTensor()
unpooler.output = torch.FloatTensor()

local testunpool = function(inputs)
   local pooled = pooler(inputs.W1)
   local unpooled = unpooler(pooled)
   return torch.sum(unpooled)
end

print(testunpool({W1=W1}))
df = grad(testunpool)
print(df({W1=W1}))

If you are using float tensors, it's better to cast the modules to :float() instead of having to manually hack it

@fmassa, Sorry, I'm not a torch expert. Can you give me an example as to what you mean?

I think he means pooler:float() instead of assigning float types to its member tensors, like pooler.output = torch.FloatTensor()